Capturing changes in flood risk with Bayesian approaches for flood

damage assessment

Flood risk is a function of hazard as well as of exposure and vulnerability. All three

components are under change over space and time and have to be considered for

reliable damage estimations and risk analyses, since this is the basis for an efficient,

adaptable risk management. Hitherto, models for estimating flood damage are

comparatively simple and cannot sufficiently account for changing conditions. The

Bayesian network approach allows for a multivariate modelling of complex systems

without relying on expert knowledge about physical constraints. In a Bayesian

network each model component is considered to be a random variable. The way of

interactions between those variables can be learned from observations or be defined

by expert knowledge. Even a combination of both is possible. Moreover, the

probabilistic framework captures uncertainties related to the prediction and provides

a probability distribution for the damage instead of a point estimate. The graphical

representation of Bayesian networks helps to study the change of probabilities for

changing circumstances and may thus simplify the communication between

scientists and public authorities. In the framework of the DFG$-$Research Training

Group "NatRiskChange" we aim to develop Bayesian networks for flood damage and

vulnerability assessments of residential buildings and companies under changing

conditions. A Bayesian network learned from data, collected over the last $15$ years in

flooded regions in the Elbe and Danube catchments $($Germany$),$ reveals the impact

of many variables like building characteristics, precaution and warning situation on

flood damage to residential buildings. While the handling of incomplete and hybrid

$($discrete mixed with continuous$)$ data are the most challenging issues in the study on

residential buildings, a similar study, that focuses on the vulnerability of small to

medium sized companies, bears new challenges. Relying on a much smaller data

set for the determination of the model parameters, overly complex models should be

avoided. A so$-$called Markov Blanket approach aims at the identification of the most

relevant factors and constructs a Bayesian network based on those findings. With

our approach we want to exploit a major advantage of Bayesian networks which is

their ability to consider dependencies not only pairwise, but to capture the joint

effects and interactions of driving forces. Hence, the flood damage network does not

only show the impact of precaution on the building damage separately, but also

reveals the mutual effects of precaution and the quality of warning for a variety of

flood settings. Thus, it allows for a consideration of changing conditions and different

courses of action and forms a novel and valuable tool for decision support.

Questions:

$1.$ What are the factors that needs to considered in assessing the flood situation

based on the climatical conditions and geographical areas?

$2.$ What are the advantages of using Bayesian network approach